Computer Networks 94 (2016) 176188

Contents lists available at ScienceDirect

Computer Networks
journal homepage: www.elsevier.com/locate/comnet

Distributed hierarchical game-based algorithm for downlink

power allocation in OFDMA femtocell networks
Song Han a, Xinbin Li a,, Zhixin Liu a, Xinping Guan a,b
a
b

Institute of Electrical Engineering, Yanshan University, Qinhuangdao 066004, China

School of Electronic Information and Electrical Engineering, Shanghai Jiaotong University, Shanghai 200030, China

a r t i c l e

i n f o

Article history:
Received 7 May 2015
Revised 15 September 2015
Accepted 10 November 2015
Available online 19 November 2015
Keywords:
Femtocell network
OFDMA
Game theory
Power allocation

a b s t r a c t
Femtocell is a promising technique for mobile operators to improve coverage and to provide
high-data rate services in a cost-ecient manner. In this study, we investigate the problem of
downlink power allocation in an orthogonal frequency division multiple access femtocell network. Femto-access points and macro-base stations in the network maximize their capacity
and compete with each other through power allocation. This competition is captured in the
framework of a Stackelberg game, and the Stackelberg equilibrium is introduced as the problem solution. We divide the strategy of MBSs into two substrategies (channel selection strategy
and power strategy) to improve the performance of MBSs. The two substrategies interact with
each other in the process of achieving Stackelberg equilibrium. The hierarchical iterative algorithm is proposed for MBSs and FAPs to achieve the maximum capacity game equilibrium in a
distributed manner. A dynamic step size mechanism and a method that copes with variational
player number are proposed to decrease the algorithm cost. Numerical results show that the
performance of the proposed algorithm is better than the performances of other algorithms.
2015 Elsevier B.V. All rights reserved.

1. Introduction
As a result of the widespread use of wireless devices and
the rapid diversication of wireless services, data trac in
wireless communication networks is explosively increasing.
More than 50% of voice services and 70% of data trac have
occurred indoors in recent years [1]. Thus, insucient indoor
coverage and the demand for high data rates are some of the
problems in existing cellular systems. Femtocell emerges as
one of the most promising technologies for improving indoor wireless service. Femto-access points (FAPs) are lowcost, low-power, and short-range base stations installed by
home users to improve indoor voice and data reception [2,3].

Corresponding author. Tel.: +86 13582406038.

E-mail addresses: hansongysu@sina.cn (S. Han), lixb@ysu.edu.cn (X. Li),
lzxauto@ysu.edu.cn (Z. Liu), xpguan@sjtu.edu.cn (X. Guan).
http://dx.doi.org/10.1016/j.comnet.2015.11.011
1389-1286/ 2015 Elsevier B.V. All rights reserved.

A variety of technical challenges in mass deployment of

femtocells have been addressed in recent years. Given that
femtocells share the same licensed frequency spectrum with
macrocells, co-channel interference affects the overall performance of femtocell networks, and the capacity of macrobase stations (MBSs) and FAPs is considered one of the major
challenges in femtocell networks [4].
Schemes based on game theory for power allocation in
wireless networks have been widely investigated in recent
years [57]. Most works have dealt with the Nash equilibrium (NE) of the resulting power control game [810], and
a few works have modeled the problem as a Stackelberg
game. In [9,10], the multiple-user multiple-channel power
allocation problem was considered, and the NE of the resulting game in orthogonal frequency division multiple access (OFDMA) femtocell networks was addressed. However,
in the algorithms, subchannels were allocated rst, followed
by power, and the change in the power strategy did not affect the subchannel strategy, thereby leading to non-optimal

S. Han et al. / Computer Networks 94 (2016) 176188

results. In [11], a Nash game was applied to copy with the

downlink interference management in a femtocell network
assuming that the MBSs power was constant, and without
changing the MBSs strategy. Femtocell networks are twotier, and the hierarchical Stackelberg equilibrium (SE) performs better than the one-tier NE. Accordingly, most works
[1216] have been conducted using the Stackelberg game
in the power control problem, including pricing data and
power. In [15], MBS behaved as a leader and controlled the
cross-tier by setting the prices for interference to each FAP,
while FAP behaved as a follower that responded to the price
strategy by optimizing the power allocation strategy; the authors proposed an ecient heuristic price-searching mechanism in the game. In [16], a Stackelberg-game-based power
control was formulated to maximize the capacity under
cross-tier interference constraints without considering subchannel allocation. Pricing mechanisms are useful in developing decentralized optimal power allocation for wireless
networks without interference. However, methods do not result in globally optimal power allocations for cellular networks with interference [17]. In [18], the authors proposed a
concept in which SE was considered a solution to the power
control in wireless networks, and they showed that SE was
superior to NE. However, only one-leader one-follower game
with two subchannels was analyzed. In [19,20], the authors
considered the interference management for femtocell networks, and the analytical solution was obtained for the Stackelberg game with only one subchannel. The work of Ngo et al.
[21] developed an ecient iterative algorithm that assigned
subchannels to receivers, allocated power to transmitters,
and alternatively searched for the optimal joint allocation solution. In [22], the author proposed an ecient algorithm for
a multiple-leader multiple-follower Stackelberg game and
proved that SE was better than NE in downlink power allocation in OFDMA femtocell networks. Nevertheless, the leaders
needed to have perfect knowledge of all channel gains and
know the channel selection strategy of all the corresponding
followers. Thus, considerable information exchange caused
the algorithm to become high cost and undistributed.
In this study, we address the issue of transmit downlink
power control with multiple channels and users by each station in interference channels with a total power constraint
base on OFDMA. We aim to devise a distributed low-cost and
high-throughput joint subchannel and power allocation algorithm for hierarchical networks. We apply the distributed
downlink power allocation method. The station allocates its
power to multiple subchannels without the knowledge of
channel gains and the channel selection strategy of others
to improve the capacity in OFDMA femtocell networks. We
assume that every station has an objective to maximize its
individual capacity. Thus, the hierarchical downlink power
control problem can be formulated as a Stackelberg game in
which MBSs are the leaders and FAPs are the followers. The
player in the Stackelberg game competes with other players
to maximize its utility. The solution of such a game is SE.
We propose a hybrid strategy that combines channel selection and power strategies for leaders to achieve the maximum capacity SE. The leaders can abandon some low-quality
subchannels to achieve more capacity from other subchannels. In other words, the leaders maximize capacity based
on the balance between bandwidth and low interference. If

177

the leaders once create some holes in spectrum, then the

followers exploit them for more capacity. Our algorithm efciently improves the capacity of both MBSs and FAPs with
low-cost mechanisms under high-interference conditions.
The main contributions of this study are as follows:
1. We propose a distributed algorithm for a multipleleader multiple-follower Stackelberg game. The solution (SE) eciently improves the capacity of a femtocell network.
2. We divide the multiple-subchannel power allocation
strategy into two substrategies, namely, the channel
selection strategy and the power strategy. The two
substrategies interact with each other in the game to
achieve SE rather than x one and then search the
other one.
3. A subchannel quality-estimating method is proposed.
MBSs estimate a subchannel based on the method and
then generate the channel selection strategy.
4. A mechanism for dynamic step size is proposed to accelerate the algorithm convergence. The iterative time
and the algorithm cost are decreased.
5. Our proposed algorithm can cope with a variational
player number at a minimal cost.
The rest of this paper is organized as follows. In Section 2,
the system model for downlink transmission is presented
and the Stackelberg game model is formulated. In Section 3,
the problem formulation and solution for sub-games are presented. In Section 4, the detailed algorithm for Stackelberg
equilibrium is presented and analyzed. Then, Section 5 shows
the simulation results and analysis of them. Finally, conclusions are given in Section 6.
2. System model
2.1. Hierarchical OFDMA network model
An OFDMA system is composed of macrotier and femtotier. As shown in Fig. 1, the transmitters of each tier are
made up of MBSs and FAPs. Let M = {1, . . . , M} denote the
set of MBSs and N = {1, . . . , N} denote the set of FAPs. Spectrum is divided into L subchannels indexed by l, with l L =
{1, . . . , L}. The set of orthogonal subchannels is shared by and
universally reused in both tiers. Each subchannel has a unit
bandwidth.
is denoted as the link gain for the channel l of transmitglk
ij
ter i to the channel k of receiver j . We assume that a channel
state is stationary in one time slot. Given the orthogonality
= 0 for k = l. The link gains are norof subchannels, gain glk
ij

=
1 for all i M N and k L. Noise
malizing such that gkk
ii
is additive, white and Gaussian, with power nkj in the subchannel k of receiver j. The power of transmitter i in subchannel k is denoted as pki . The interference from other transmitters observed by the receiver j in subchannel k is given by

vkj = i= j gkij pki + nkj .
We assume that the receivers can distinguish the useful
signal from the interference and know the power value. The
signal to interference plus noise power ratio is observed by
the receiver j in channel k is given by

178

S. Han et al. / Computer Networks 94 (2016) 176188

Fig. 1. Topology of the two-tier femtocell network.

jk = 

pkj
k k
i= j gi j pi

nkj

The transmitter and the receiver comes in pairs. For brevity,

we denote vi as the interference of receiver j corresponding
to transmitter i, so does i .
The set of the transmit power vector of transmitter i is
given by Pi = [p1i , p2i , . . . , pLi ]T , and the power of transmit is

.
subject to a total power constraint kL pki pmax
i
Similarly, the interference vector of receiver i is Vi =
[v1i , v2i , . . . , vLi ]T . And the set of the transmit power vector of
transmitters is given by P = [P1 , P2 , . . . , PM , PM+1 , . . . , PM+N ] .
We consider the distributed power allocation problem
with the total power constraint, in which each transmitter allocates its power among L subchannels to maximize its total
capacity. Each receiver considers the interference from other
transmitters as an additive white Gaussian noise. If power allocation is given, then the data rate of user i is provided by
the Shannon formula

Ci =


kL

log2 1 +

pki

vki

2.2. Stackelberg game model

The Stackelberg game is a strategic game that consists of
leaders and followers who are competing with each other
for a certain resource. The leaders in the game have a higher
level than the followers do. The leaders have the privilege of
making the rst move, and the followers move after the leaders. Therefore, a distinct hierarchy exists between the players. The leaders can anticipate and take the behavior of the
followers into consideration before making their own move.
The followers do not have the ability to anticipate the move
of the leaders; thus, they can only react to the leaders.
In a hierarchical OFDMA femtocell network, macrotier has
a higher level than femtotier. Therefore, we formulate MBSs

as the leaders that belong to the upper subgame and FAPs as

the followers that belong to the lower subgame. The set of

players in the Stackelberg game is thus M N.
The multiple subchannel power allocation problem consists of two subproblems, namely, Which subchannels are
selected to allocate power? (channel selection) and How
much power should be allocated over each selected subchannels?(power allocation). In a wireless network, power allocation is tightly coupled with channel selection [23]. The
joint subchannel and power allocation issue is non-trivial because of the intrinsic non-convexity when dealing with the
combinatorial nature of multiple-carrier systems [24].
Thus, we formulate the multiple subchannel power allocation strategy as a hybrid strategy in the Stackelberg game.
The strategy space of each player i is its power allocation
Pi = [p1i , p2i , . . . , pLi ]T . pki = 0 represents that the user i does
not use the subchannel k, i.e., subchannel k is inactive; otherwise, it is active. User i abandons the subchannel k (inactive) because of two reasons. One reason is that user i cannot achieve prot capacity in subchannel k, i.e., user i abandons subchannel k passively. The other reason is that user i
abandons subchannel k voluntarily to generate more capacity on other channels. Femtocells are expected to exploit the
leftover resource opportunistically. Therefore, only MBSs
can abandon channel voluntarily. In this study, we divide the
strategy of leader Pi , i M into two spaces (hybrid strategy),
namely, channel selection strategy SMi and power strategy
PMi . The two strategies interact with each other in the game.
And SMi = [s1Mi , s2Mi , . . . , sLMi ],


skMi =

1, i f the user i not abandon the subchannel k (active )

0, otherwise (inactive ).

The strategy of follower i is power allocation PNi , i

N. The channel selection state of follower i is thus SNi =

S. Han et al. / Computer Networks 94 (2016) 176188

[s1Ni , s2Ni , . . . , sLNi ],

1,
0,

skNi =

3. Problem formulation and solution for subgames

i f pkNi > 0
otherwise

(active )
(inactive ).

Let Ali and A denote the sets of active subchannels of the


leader and the follower, respectively. Hence, |Ali | = SMi ,

and |A f i | = SNi . || represents the cardinality of a set.
In this study, the objective of each player is to maximize
its capacity by adjusting its strategy. Players compete and interact with each other.
Denition 1. The lower (upper) subgame strategy

(P low (P up )) is xed. Point P up (P low ) is the upper (lower)

subgame NE if for all players in the upper (lower) subgame i
M (i N), the following conditions are satised:

up
up
Ci (Piup , Pi
, P low ) Ci (Piup , Pi
, P low ),

low
Ci Pilow , Pi
, P up

( (

179

) (

low
Ci Pilow , Pi
, P up

max

Pi

1+ 

log2

s.t. pki 0,

L


k k
j=i g ji p j

(6)

Existence of SE:
Proposition 1. [25] An NE exists in game G p,k = [M, pki , Cik ]. If
for all i M , then the following statements hold.
(1) pki is a non-empty convex and compact subset of some
Euclidean space RK .
(2) Cik ( pki ) is continuous and concave in pki .
Theorem 1. An NE exists in the subgame (upper and lower
games)G p,k = [M (N ), pki , Cik ]
Proof. The strategy space is dened as pki : 0 pki pmax
,
i
and it is a non-empty, convex, and compact subset of the Euclidean space RK .
Let



pki
Cik log2 1 +  j=i gkji pkj +nki
1
=
= 2
.
pki
pki
ln (vki + pki )

Pj

(2)

(3)

(4)

Cik ( pki ) is concave in ( pki ). According to Proposition 1, both

the upper and the lower games have NE. 
Corollary 1. Assuming that the best response function of the
upper (lower) game is continuous, at least one SE, composed by
the NEs of subgames, exists for the hierarchical game.
Proof. Considering that the best response functions BRup
and BRlow are continuous functions, the composition of these
two functions BR is also continuous. Pi is convex, closed and
bounded. Given that the SE is dened as any xed point of
BR() , from the Schauder xed point theorem that every continuous function from a convex compact subset of a Banach
space to itself has a xed point [26], the existence of SE is
proven. 

(5)

+ nki

pki =pmax
i

Pj argmax C j : pkj 0,

)),

pki

k=1

(1)

where Pi is a vector of the strategies of all players, except

player i.
In the two-tier Stackelberg game, the SE is constituted by
two subgame equilibriums. Therefore, SE is dened as P =

(Pup , Plow ).

2Cik
1
= 2
< 0.
k 2
k
( pi )
ln (vi + pki )2

In this study, the solution of the Stackelberg game can be

achieved by the NEs of two subgames. Therefore, power allocation is divided into two parts to be analyzed and solved.
In the game, each player rst observes the inuence from
other players and then changes its strategy to achieve more

capacity. The best strategy of each player i M N can thus
be found by the following optimization problem with equilibrium constraints:

L


j M N.

pkj = pmax
,
j

k=1

(7)
We assume that all users can sense and differentiate a
useful signal from an interference signal. For description convenience, a single-leader and multiple-follower scenario is
formulated. To ease the notations, we let PM , SM , and Al replace PMi , SMi , and Ali , respectively.
We make the same assumption as in [22], i.e., subgames
converge to a unique and stable NE during the iterative waterlling (IW) process. The spectral radius of the channel gain
matrix should be less than unity for the stability and convergence of the IW process [27].
3.1. The solution of the lower game
Problem 1 : The lower game for the follower i is

max CNi = max

PNi

PNi

L


log2 1 +

k=1

gkli pkM +

pkNi
k k
j=i g ji pN j

+ nkNi

(8)

k 0. Here, gk is the link
such that Lk=1 pkNi = pmax
and
p
Ni
Ni
li
gain from transmitter of the leader to the receiver of follower i on subchannel k. The noise-plus-interference term,

vkNi = gkli pkM + j=i gkji pkN j + nkNi , is treated as a constant by
the follower. The objective function is concave. Therefore, the
well-known IW algorithm can be the solution of the lower
game [27], i.e.,

PNi =

0
Ki vkNi

f or k
/ Afi
f or k A f i ,

(9)

where 1/Ki is the positive Lagrange multiplier. By the con

( pmax
+ kA vkNi )

Ni
fi
,
we
can
have
K
=
.
straint Lk=1 pkNi =pmax
i
Ni
|A |
fi

The best response of the followers can be generated by

iterating Eq. (9) until convergence is achieved if the strategy
of the leader is constant.
3.2. The solution of the upper game
Problem 2: The upper game for the leader is

max CM = max
PM

PM

L

k=1

log2

1+ 

pkM
k k
iN hi pNi

+ nkM


,

(10)

180

S. Han et al. / Computer Networks 94 (2016) 176188


such that Lk=1 pkM = pmax
and pkM 0. Here, hki is the link
M
gain from the follower i transmitter to the leader receiver on
subchannel k. The solution is similar to the solution of the
lower game i.e.,

PM =

0
W vkM

f or k
/ Al
f or k Al ,

(11)

where 1/W is the positive Lagrange multiplier. By the con

( pmax
+ kA vkM )

M
l
straint Lk=1 pkM =pmax
,
we
can
have
W
=
.
M
|A |
l

The interference of each user on each subchannel is the

main information for the IW algorithm. We dene the system
interference vector by V.
T

V = [V1T , . . . , VNT VMT ] .

(12)

The system interference vector can be obtained by the

system power by:

V = Gp + n,

(13)

where n is the column vector expressed as follows:

n=

Fig. 2. Algorithm for SE.

T
[n1N1 , . . . , nLN1 , n1N2 , . . . , nLN2 , n1NN , . . . , nLNN , n1M , . . . , nLM ] .

(14)
Matrix G can be interpreted as the gain matrix with gains
and zeros as entries. For example, when M = 1, N = 2, and
L = 2, the interference vector V is

1
0
vN
1
v2 0
N1
1 11
vN 2 g12

v 2 = g 21
N 2 12
1 11
vM h1
2
vM
h21
1

0
0
12
g12
22
g12
h12
1
h22
1

11
g21
21
g21
0
0
h11
2
h21
2

12
g21
22
g21
0
0
h12
2
h22
2

11
gl1
21
gl1
11
gl2
21
gl2
0
0

1 1
12
gl1
nN 1
pN 1

22 2
gl1
pN 1 n2N 1
1

12 1

gl2
p
n
N2 + N2 .
22 2
n2
gl2
p
N2 N2

0 p1M n1M
2
2
0
pM
nM

can obtain the best power strategy of the leader by iterating

Eq. (11).
Thus, the key to obtain the solution of the Stackelberg
game is the channel selection strategy SM (or Al ). We describe
the algorithm for SM (or Al ) and the equilibrium of the Stackelberg game in Section 4.
4. Algorithm for the optimal strategy in the Stackelberg
game

(15)

Matrix G is a partitioned matrix with zero diagonal blocks. Eq. (15) is the matrix expression of vkNi =


gkli pkM + j=i gkji pkN j + nkNi , f or Ni N and vkM = iN hki pkNi +
nkM , f or M M. Therefore, the interference of each user (in
both of the lower game and the upper game) on each subchannel can be found in V.
The game players do not need detailed knowledge of the
strategies of other players or the link gains. All they need is
the ability to know the power of useful signal and the interference. They can directly self-adjust their strategies by sensing the interference vki on their own subchannels. As a result, they can generate strategies in a distributed manner and
avoid learning high-dimensional information in Eq. (15). The
information of gains and the strategies of followers grow exponentially with the number of subchannels (or users) from
others.
The difference between the two subgames is the channel
selection. From the analysis of Section 2.2, we know that the
leader can abandon the bad subchannel voluntarily, even if
the subchannel is a positive prot for it. By contrast, given
that the followers cannot do that, they can only utilize the
leftover resource. If they can generate a positive prot from
the subchannel, then they take it or leave it.
Therefore, if both the strategy of the followers and the
channel selection strategy of the leader are constant, then we

The algorithm for SE is rst shown, followed by details

about the algorithm for the hybrid strategy of the leader.
4.1. Algorithm for SE
The hierarchical Stackelberg game is constituted by two
subgames. In the Stackelberg game, a follower has to change
its best response to adapt to the change from other players,
the leader, and other followers. If the leader nds that the
strategy of a follower has changed, then it must adjust its
strategy to enhance utility. This phenomenon naturally leads
to an iterative method of nding SE, as shown in Fig. 2.
Fig. 2 shows the framework of the algorithm for SE. The
part about the solution to the subgame of the followers is
shown. The remaining work is to nd the calculation method
for the best power strategy and the channel selection strategy
(the hybrid strategy of the leader), which is the key of the
algorithm.
4.2. Algorithm for the hybrid strategy of the leader and the
relevant optimal mechanisms
4.2.1. Algorithm for the hybrid strategy of the leader
The total transmit power and the number of subchannels
are constant. The essence of the problem is how to allocate
the limited power to the limited subchannel to maximize
utility. The diculty of solving the problem is that the strategy is not single but hybrid and distributed computing. In
other words, the key problem is how to compute the bestchannel selection strategy and power strategy by using a distributed algorithm.

S. Han et al. / Computer Networks 94 (2016) 176188

For the leader (MBS), the worth of different subchannels

differs. However, the total transmit power is constant, and
the leader prefers high-quality subchannels. Thus, the leader
can abandon some bad subchannels voluntarily to achieve
more capacity on good subchannels. The essence of the abandon strategy is whether the increment capacity of good subchannels is larger than the reduction capacity of abandoned
subchannels. If the increment capacity of the unabandoned
subchannels is larger, then the abandon strategy is benecial. Otherwise, it is inadvisable. Thus, the leader can balance
the bandwidth and low interference to maximize capacity by
using the abandon strategy.
The capacity of the leader on subchannel k is as follows:

k
CM

= log2 1 +

pk

vk

= log2 1 +

W vk

vk

= log2

W
vk

(16)
and



k
log2 1 +  hpk Mpk +nk
CMk
iN i Ni
M
=
pkM
pkM
=
=
=

ln

1
k k
k
k
iN hi pNi + nM + pM

1
ln (vkM + pkM )

vkM

is the step size of the channel selection strategy searching, and we assume that = 1 to simplify the instruction.
The dynamic step size mechanism is explained in the ensuing
paragraphs.
Fig. 3 shows the detailed process of the hybrid strategy.
The left part is performed by the leader, and the right part is
performed by the followers. Initialize means the power allocation of SE without the strategy that the leader voluntarily
abandons a subchannel and T = 0. First, the leader estimates
the quality of all subchannels, and then obtains Q. Second,
the leader arrange Q in ascending order; thus, the leader can
know which subchannel quality is low. Third, the leader tentatively abandons subchannel that has the minimum quality. The leader then observes SE and estimates whether its
utility increases. If the utility increases, then the abandon
strategy is correct, and the leader creates more holes. Otherwise, the abandon strategy is incorrect. Hence, the algorithm is ended, and PM (T 1 ), PN (T 1 ) are the best solutions for the hybrid strategy in the Stackelberg game.
The proposed algorithm is distributed. Users do not need
to know the specic strategy of one another and the channel
gain with other transmitters. They generate a strategy by observing the interference and the useful signal they received.
Users can adaptively achieve the best strategy with the proposed algorithm.

1
2

ln W

(17)

For k Al , W does not change with the subchannel. From

Eq. (17), if we invest equivalent power in every subchannel
that belongs to the leader, then the increment in capacity is
equal. In other words, the increment in capacity from the unabandoned subchannel is equal. Thus, the reduction in capacity from abandoned subchannels is the criterion for subchannel quality. Obviously, the lower-worth subchannel is the one
that has a lower capacity.
On the basis of the analysis of Eqs. (16) and (17), we propose a quantitative criterion Qk for estimating the quality of
subchannel k.

Qk =

181

, k Al .

(18)

Q = (Q 1 , Q 2 , . . . , Q k ). Qk denotes the willingness of the

leader to increase the power on subchannel k. In other words,
if Qk is larger on subchannel k, then subchannel k is more
worthy of investing more power in the subchannel. Otherwise, the subchannel is not worthy of investing power, and
the leader should abandon this subchannel and invest power
into the better subchannel to improve its utility. When the
leader abandons the inecient subchannel, holes are created for the followers. The followers exploit these holes by
allocating most of their power in these holes for more utility. The behavior of the followers decreases the interference
of the leader from the followers. If the abandonment of the
low-quality subchannel strategy can improve the utility of
the leader, then the holes also improve the utility of the
followers. Therefore, the good-channel selection strategy can
improve the utility of the entire femtocell network.
The algorithm for the hybrid strategy is presented in
Fig. 3.

4.2.2. Dynamic step size mechanism

Given that the game is hierarchical and the strategy of the
leader is hybrid, the algorithm needs to be a nested loop to
reach SE. Thus, the iterative times of the outer ring have a
great inuence on the algorithm speed. We propose a mechanism for dynamic step size to accelerate the convergence as
follows:
Mechanism for dynamic step size
calculate I = T 1
if T = 1
set (I ) = 0, = 1, f l = 1
end if
if (CM (T ) > CM (T 1 ) & f l=1)
set su = 1
else
set su = 0, f l=0
end if
if su = 1
if T = 2
calculate (I ) = CM (I )CM (I1)
end if
if T > 2
calculate (I ) = CM (I )CM (I1)
set Fo = (I )

if (I ) (I 1 )
set = + Fo
else
set = max( Fo, 2 )
end if
end if
else
set = 1
the leader searches the best channel selection strategy
around SM (T 1 ) by the step size = 1
26:
the leader can reuse the highest-quality subchannel kL
kL
= 1) or abandon
in the abandoned subchannels (set SM
k1
= 0) tentatively
the lowest-quality subchannel k1 (set SM
27:
until achieve the maximum utility
28: end if

1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:

182

S. Han et al. / Computer Networks 94 (2016) 176188

Fig. 3. Algorithm for the hybrid strategy.

The essence of the mechanism is that the leader adjusts

its search step size based on the variation trend of its utility.
If the strategy of abandoning subchannels increases (I)
augmenter of the utility per subchannel, then the step size
should increase = + Fo. Otherwise, the step size should
decrease; thus, = max( Fo, 2 ).
The termination condition of the algorithm with a dynamic step size differs from that of the algorithm with a xed
step size = 1. This difference is due to the fact that if > 2,
the rst time in CM (T ) < CM (T 1 ), CM (T 1 ) may not be
the maximum utility. The best point is around SM (T 1 ),
which we call it nearby point. When we have the nearby
point, we set su = 0 to mark the stage. After the nearby point
is found, we set = 1. The leader then searches the best
point by exploring the points around the nearby point. In
the dynamic step size mechanism, the last judgment in Fig. 3
should be replaced with (su = 1||CM (T ) > CM (T 1 )).
4.3. Mechanism for the countermeasure of variational players
In an actual femtocell network, the number of players constantly changes. Considerable works have been conducted on the resource allocation for femtocell networks, but
limited works have considered the countermeasure for variational players. Most of the algorithms request that the number of users be xed. If the number of users changes, then the
algorithm should restart, which leads to frequent changes in
the power of users. Therefore, the algorithms become inecient and have a low performance of users in practice.

Given that the proposed algorithm in this study is distributed, the players in the game adjust their strategies based
on sensory information (the interference and the useful signal). They do not require knowing the details of the interference source and the strategies of other players. Hence, they
can cope with the variational players easily. The mechanism
for the countermeasure of variational players is as follows:
If the leader senses that the variation in interference
exceeds a value (which is stated by the system), then it
will abandon the lowest-worth subchannel tentatively. If the
abandon strategy can improve its utility, then the leader
applies the abandon strategy until the maximum utility is
achieved, or it reuses the highest-worth subchannel that is
not occupied until the maximum utility is achieved.
The initial value of the original players should therefore
be the strategy in the previous stage. The power does not
need to be initialized to zero or an average value. Users can
adaptively adjust their strategy to achieve the maximum utility. The mechanism makes history information useful to decrease the strategy change frequency of users.
4.4. Convergence of the algorithm
Proposition 2. The lower game is convergent if the upper game
strategy is xed.
Proof. In the lower game, the IW algorithm is used to
achieve NE with xed power allocation of the leader. Thus,
in the process of achieving the lower game NE, the followers

S. Han et al. / Computer Networks 94 (2016) 176188

takes the power of the leader as a xed interference. The

strategy of the leader does not affect the convergence of the
followers. In our case, the sucient convergence of the IW
algorithm is assumed as in [22]. 
Proposition 3. The upper game is convergent if the lower game
strategy is xed.
Proof. In the upper game, the IW algorithm is used under
the tentative channel selection. The strategy in which the
leader abandons some worthless subchannels decreases the
spectral radius of the channel gain matrix. Thus, the abandon
strategy makes the convergent condition of the IW algorithm
easier. 
Proposition 4. The algorithm for the hybrid strategy in the
Stackelberg game is convergent.
Proof. From Corollary 1, at least one existing lower NE corresponds to an upper NE and vice versa. From Propositions 1
and 2, the upper-lower game is convergent if the subgame NE
exists. Thus, the above statement is proven. 
Discussion. The key to the success of the proposed algorithm lies in two points.
1. One point is the ability of the user to know the power
of useful signal and the interference that is similar
to cognitive radio. The ability ensures that the algorithm can be used in a distributed mode. Unlike
undistributed algorithms, the proposed distributed algorithm allows users adaptively search the best strategy by self-sensing the interference without the assumption that leaders must have perfect knowledge of
all channel gains and the channel strategies of the followers, such as in [22]. The advantage of a distributed
algorithm is not only the low channel cost but also
the decreased spectral management cost. If the leaders
have to know a specic strategy from the report of the
followers, then the followers may report the dishonest information for more utility. Thus, a mechanism for
honesty must be built against the dishonest strategy
with a high spectral management cost.
2. The other point is the high co-channel gain. When the
co-channel gain is high, the leader avoids causing high
interference and voluntarily abandons the worthless
subchannels that have high interference, thus creating
spectral holes. For the followers with low total transmit power, spectral holes are valuable. The followers
allocate more power in spectral holes, thus decreasing the interference over other subchannels. In the
entire process, the leaders lose bandwidth, but they
obtain higher-quality and lower-interference subchannels. In [22], the strategy of the leader totally depends on the channel strategy of the followers, without the mechanism of voluntary channel selection for
the leader. Consequently, the utility of the leader is
not maximized. Unlike in [22], in this study, when the
leader with the voluntary channel selection mechanism achieves the maximum utility under SE, the algorithm ends. The leader has the absolute priority.
Hence, this criterion is satised by femtocell networks
whether under a high interference or MBS priority.

183

Extension to multiple leaders. In the multiple-leader scenario, the interferences from the strategies of other leaders
are regarded as the background noise of each leader. The
leaders iterate their best response among themselves until
the upper subgame NE is reached. The followers use the waterlling algorithm to reach the lower subgame NE. Thus, the
process of searching the NE of the two subgames is iterated
until the entire process converges.
Scalability analysis of the algorithm. In the game, we offer the upper-tier users (MBS) a more exible strategy (hybrid strategy) to maintain high-level priority for users. Thus,
the algorithm can be used in other hierarchical networks.
The proposed algorithm is a fully distributed model. The algorithm can be easily extended to any other multiple-user
multiple-subchannel wireless networks, especially in highinterference conditions. Considering that we add the channel selection strategy based on the IW algorithm, the proposed algorithm performance cannot be lower than the IW
algorithm even in friendly interference conditions.
5. Numerical results
In this section, we present the simulation results to evaluate the performance of the proposed algorithm in comparison with the performances of the IW algorithm and the algorithm in [22].
A single-leader and two-follower scenario is considered
in the simulations. The amount of subchannels L = 100, and
the total power of a FAP, follower, is taken to be 20 dB bemax =
low the total power of MBS, leader. We assume that PM
30 dBm, and PNmax = 10 dBm. The noise power in each subchannel is WGN with a power of 120 dBm. Each subchannel is assumed to undergo independent and identically distributed Rayleigh fading. For each subchannel, the Rayleigh
random variable is modeled as the modulus of circularly
symmetrical complex Gaussian (CSCG) distributed. The average gain of a subchannel is controlled by changing the variance of the CSCG distribution. We set the iterative precision
105 . To verify the universal validity of the algorithm, we
base each plot on 100 random channel gain samples and observe the average performance, as shown in Table 1.
Table 1 shows the capacity (utility) of users with different algorithms. We are particularly interested in examining
the performance under high-interference conditions. The average direct channel gain is maintained to be unity, while the
average co-channel gain is kept to be 0.4 and 0.5. The second,
third, and fourth columns show the average capacity of the
leader, the followers, and the entire system, respectively.
From Table 1, the capacity of both the leader and the followers at SE is improved in contrast to NE. The capacity of
the leader is higher in a higher-gain condition. The leader
has more power than the follower, and the preponderance
is more obvious in a higher-gain condition. When the cochannel gain is higher, then the average co-channel gain is
0.5, and the algorithm in [22] makes the followers superior. The capacity of the followers is 97.3750, which is larger
than the capacity with the other algorithms. The leader creates signicant holes, thereby leading to the higher capacity of the followers, without increasing the capacity of the
leader. However, the capacity of the leader is two orders of

184

S. Han et al. / Computer Networks 94 (2016) 176188

Table 1
Comparison with other algorithms.
Avg. co-ch. gain 0.4

Capacity of the leader

Capacity of the followers

Capacity of the system

IW (NE)
[22] (SE)
This paper (SE)
Avg. co-ch. gain 0.5
IW (NE)
[22] (SE)
This paper (SE)

1538.2
1582.7
1694.6
Capacity of the leader
1551.3
1558.7
1710.1

23.2570
38.2557
42.4782
Capacity of the followers
20.3323
97.3750
38.8158

1561.46
1620.96
1737.08
Capacity of the system
1571.60
1656.08
1748.92

1.4
leader
followers

1.2

CDF

0.8

0.6

0.4

0.2

1.2

1.4

1.6

1.8
2
C /C
SE

2.2

2.4

2.6

2.8

NE

Fig. 4. Ratio of capacities at Stackelberg equilibrium using our algorithm and Nash equilibrium when average co-channel gain is 0.4.

magnitude higher than that of the followers. The leader is a

more important part of the system accordingly. Owing to the
leader priority and the rational termination condition, our
algorithm performs better in the capacity of the leader and
the entire system. Our algorithm also shows higher performance in the higher-interference condition. Hence, our algorithm is t for an actual femtocell network with a high
interference.
We plot the cumulative distribution function (CDF) of
the ratio between the capacity obtained under our algorithm and others to compare the performances of the
algorithms. Figs. 47 show the CDF under average cochannel gains of 0.4 and 0.5, respectively. As mentioned
above, the capacity of the leader is higher than that of the
followers, and thus the curve of the entire system is similar
to that of the leader. Therefore, the curve of the entire system
is not given.
Figs. 4 and 6 show the ratio of capacity at SE using our
algorithm and the NE using the IW algorithm under different co-channel gains. Both gures show that the performance
of our algorithm is generally better than that of the IW algorithm. Whether the average co-channel gain is 0.4 or 0.5,
no point lower than the IW algorithm exists. Most of the ratios of the capacity of the followers are between 1.5 and 2.5.

Under different co-channel gains, Figs. 5 and 7 show the ratio of capacity at SE using our algorithm and the algorithm in
[22], respectively. When the average co-channel gain is 0.4,
approximately 70% of the ratios of the capacity of the followers obtained by our algorithm are higher than those obtained by the algorithm in [22]. When the average co-channel
gain is 0.5, the capacity of the followers obtained by the algorithm in [22] is higher than that obtained by our algorithm.
However, 96% of the ratios of the capacity of the leader (or
system capacity) obtained by our algorithm are higher than
that obtained by the algorithm in [22] under 0.4 average cochannel gain, and almost 100% of the ratios of the capacity
of the leader (or system capacity) obtained by our algorithm
are higher than that obtained by the algorithm in [22] under 0.5 average co-channel gain. The reason for this result is
that our algorithm focuses on preserving the priority of the
leader. Thus, in such a case, our algorithm shows a higher
performance in vast majority of cases.
We plot the iterative process in comparison with constant
step = 1 and dynamic step when the average co-channel
gain is 0.4 to show the dynamic spectral allocation process of
the proposed algorithm and the performance of the dynamic
step mechanism. The initial strategy of players is assigned
/L, pkNi =
with equal power on every subchannel, pkM = pmax
M

S. Han et al. / Computer Networks 94 (2016) 176188

185

1.4
leader
followers
1.2

CDF

0.8

0.6

0.4

0.2

0
0.4

0.6

0.8

1.2
1.4
1.6
CSE/CSE[22]

1.8

2.2

Fig. 5. Ratio of capacities with our algorithm and the algorithm [22] when average co-channel gain is 0.4.

1.4
leader
followers
1.2

CDF

0.8

0.6

0.4

0.2

0
0.5

1.5

2
C

2.5
/C

SE

3.5

4.5

NE

Fig. 6. Ratio of capacities at Stackelberg equilibrium using our algorithm and Nash equilibrium when average co-channel gain is 0.5.

pmax
/L, f or k L. Two representative samples are shown as
Ni
follows.
Given that the capacity of the followers is not in the same
order of magnitude as that of the leader, the iterative process
curves of the followers are not given in gures. We show the
entire result in Table 2.
In Table 2, the rst two rows and the last two rows summarize the results in Samples 1 and 2, respectively.
Figs. 8 and 9 and Table 2 show that the proposed dynamic
step size mechanism largely improves the algorithm speed

with only a minimal capacity loss. The key to the success of

the dynamic step size mechanism is the ability of the leader
to change the step size by learning the network environment
(observe the variation tendency of its own utility). Therefore,
the convergence speed is improved eciently.
Let us denote the change in the channel selection function CCS = S(T1 ) S(T ), such that |CCS| means that the variational player number affects the decision of users. A high
|CCS| means a signicant negative effect. A specic condition is proposed to analyze the ability of coping with the

186

S. Han et al. / Computer Networks 94 (2016) 176188

1.4
leader
followers
1.2

CDF

0.8

0.6

0.4

0.2

0.2

0.4

0.6
0.8
CSE/CSE[22]

1.2

1.4

Fig. 7. Ratio of capacities with our algorithm and the algorithm [22] when average co-channel gain is 0.5.
Table 2
Performance of dynamic step size.
Algorithm

Leader change
strategy times

Leaders active
subchannels |Al |

Leader capacity

Follower change
strategy times

Followers active
subchannels |Af1 |, |Af2 |

Followers
capacity

=1

80

90

1672.4

308

15,10

Dynamic

45

88

1661.5

173

15,12

=1

73

91

1728.9

388

12,13

Dynamic

51

91

1728.6

292

12,12

24.5902
24.2341
28.2393
26.4867
28.4628
18.2036
25.5528
18.8200

Fig. 8. Sample 1 for the dynamic process of the proposed algorithm.

S. Han et al. / Computer Networks 94 (2016) 176188

187

Fig. 9. Sample 2 for the dynamic process of the proposed algorithm.

Table 3
Effect of variational player number.
Player

|CCS|

Capacity before
follower 3 joins the
game

Capacity after
follower 3 joins the
game

Leader
Follower 1
Follower 2
Follower 3

4
5
4
-

1723.7
22.8610
22.4656
-

1641.60
20.0209
24.4961
20.8400

variational player number. Only one leader and two followers

exist in the system. After they achieve their best SE S(Ttwo ),
the third follower joins the game. Finally, the game with the
three followers achieves the SE S(Tthree ) of the new game.
Thus, CCS = S(Ttwo ) S(Tthree ) in Table 3.
Table 3 shows the effect of changing follower number
from 2 to 3. The channel selection strategies of the players
change inconsiderably, which is good for the performance of
the communication network. Their strategies can smooth the
transition to cope with the variational player number rather
than initializing all strategies as in other algorithms.

channel selection strategy based on the change in its capacity and generates a new power strategy until its capacity
achieves the maximum under the upper game equilibrium.
The hybrid strategy of leader is more exible and diverse
for maximizing its capacity. Followers use an IW to reach
the lower game equilibrium. The spectral holes created by
leaders are the main subchannels for followers. Thus, the interference of a leader from followers decreases to some extent. The two subgames are played iteratively to achieve SE.
A subchannel quality-estimating method, the dynamic step
size mechanism, and a method that copes with variational
player number are proposed to improve the performance of
the algorithm. The simulation results show the signicant capacity gains of macrocell and femtocell with a low cost.

Acknowledgment
The authors thank the nancial support of the Natural
Science Foundation of China under Grant nos. 61172095,
61571387 and 61473247.

Supplementary material
6. Conclusion
In this study, we propose a distributed algorithm to solve
the problem of downlink power allocation in OFDMA twotier femtocell networks. The objective of each station is to
maximize its capacity by power control with the total power
constraint. A hierarchical Stackelberg game is formulated.
In the upper game, we divide the strategy of a leader into
two substrategies (hybrid strategy), namely, the channel selection strategy and the power strategy. Leaders reach the
upper subgame equilibrium with their best hybrid strategies. In the process, each leader observes its capacity after
the power strategy is performed. A leader then adjusts its

Supplementary material associated with this article can

be found, in the online version, at 10.1016/j.comnet.2015.11.
011
References
[1] D. Lpez-Prez, A. Valcarce, G. De La Roche, J. Zhang, OFDMA femtocells:
A roadmap on interference avoidance, Commun. Mag. IEEE 47 (2009)
4148.
[2] V. Chandrasekhar, J.G. Andrews, A. Gatherer, Femtocell networks: a survey, Commun. Mag. IEEE 46 (2008) 5967.
[3] H. Claussen, L.T.W. Ho, L.G. Samuel, An overview of the femtocell concept, Bell Labs Tech. J. 13 (2008) 221245.

188

S. Han et al. / Computer Networks 94 (2016) 176188

[4] W.S. Jeon, J. Kim, D.G. Jeong, Downlink radio resource partitioning with
fractional frequency reuse in femtocell networks, IEEE Trans. Veh. Technol. 63 (2014) 308321.
[5] B. Wang, Y. Wu, K.R. Liu, Game theory for cognitive radio networks: An
overview, Comput. Netw. 54 (2010) 25372561.
[6] L. Song, D. Niyato, Z. Han, E. Hossain, Game-theoretic resource allocation methods for device-to-device communication, IEEE Wirel. Commun. 21 (2014) 136144.
[7] W. Wang, A. Kwasinski, D. Niyato, Z. Han, A survey on applications
of model-free strategy learning in cognitive wireless networks, arXiv
preprint arXiv:1504.03976 (2015).
[8] H.A. Shyllon, S. Mohan, A game theory-based distributed power control algorithm for femtocells, in: Proceedings of the 2014 IEEE International Conference Advance Networks Telecommunications Systems
(ANTS), 2014, pp. 16.
[9] J. Zhao, W. Zheng, X. Wen, X. Chu, H. Zhang, Z. Lu, Game theory
based energy-aware uplink resource allocation in OFDMA femtocell
networks, Int. J. Distrib. Sens. Netw. 2014 (2014) 18.
[10] H. Zhang, X. Chu, W. Ma, W. Zheng, X. Wen, Resource allocation with interference mitigation in OFDMA femtocells for co-channel deployment,
EURASIP J. Wirel. Commun. Netw. 2012 (2012) 19.
[11] S. Shen, T.M. Lok, Dynamic power allocation for downlink interference
management in a two-tier OFDMA network, IEEE Trans. Veh. Technol.
62 (2013) 41204125.
[12] Z. Liu, L. Hao, Y. Xia, X. Guan, Price bargaining based on the Stackelberg
game in two-tier orthogonal frequency division multiple access femtocell networks, IET Commun. 9 (2014) 133145.
[13] K. Zhu, E. Hossain, D. Niyato, Pricing, spectrum sharing, and service selection in two-tier small cell networks: A hierarchical dynamic game
approach, IEEE Trans Mob. Comput. 13 (2014) 18431856.
[14] Y. Xiao, G. Bi, D. Niyato, Distributed optimization for cognitive radio
networks using Stackelberg game, in: Proceedings of the 2010 IEEE International Conference Communication Systems (ICCS), 2010, pp. 77
81.
[15] W. Wang, A. Kwasinski, Z. Han, Power allocation with Stackelberg game
in femtocell networks: A self-learning approach, in: Proceedings of the
2014 IEEE Eleventh Annual International Conference on Sensing, Communication, and Networking (SECON), 2014, pp. 354362.
[16] X. Kang, R. Zhang, M. Motani, Price-based resource allocation for
spectrum-sharing femtocell networks: A Stackelberg game approach,
IEEE J. Sel. Areas Commun. 30 (2012) 538549.
[17] S. Sharma, D. Teneketzis, A game-theoretic approach to decentralized
optimal power allocation for cellular networks, Telecommun. Syst. 47
(2011) 6580.
[18] M. Bennis, M.L. Treust, S. Lasaulce, M. Debbah, J. Lilleberg, Spectrum
sharing games on the interference channel, in: Proceedings of the 2009
International Conference on Game Theory for Networks, 2009, pp. 515
522.
[19] Q. Han, K. Ma, X. Wang, X. Guan, J. Ma, Stackelberg game based interference management for two-tier femtocell networks, Wirel. Netw. 19
(2013) 16651677.
[20] Q. Han, B. Yang, X. Wang, K. Ma, C. Chen, X. Guan, Hierarchical-gamebased uplink power control in femtocell networks, IEEE Trans. Veh.
Technol. 63 (2014) 28192835.
[21] D.T. Ngo, S. Khakurel, T. Le-Ngoc, Joint subchannel assignment and
power allocation for OFDMA femtocell networks, IEEE Trans. Wirel.
Commun. 13 (2014) 342355.
[22] S. Guruacharya, D. Niyato, D.I. Kim, E. Hossain, Hierarchical competition for downlink power allocation in OFDMA femtocell networks, IEEE
Trans. Wirel. Commun. 12 (2013) 15431553.
[23] K. Son, S. Lee, Y. Yi, S. Chong, REFIM: A practical interference management in heterogeneous wireless access networks, IEEE J. Sel. Areas
Commun. 29 (2011) 12601272.
[24] H. Wang, Z. Ding, Power control and resource allocation for outage balancing in femtocell networks, IEEE Trans. Wirel. Commun. 14 (2015)
20432057.
[25] G. Owen, Game Theory, Academic, New York, NY, USA, 2005.
[26] K. Goebel, W.A. Kirk, Topics in Metric Fixed Point Theory, Cambridge
University Press, 1990.
[27] K.W. Shum, K.K. Leung, C.W. Sung, Convergence of iterative waterlling
algorithm for Gaussian interference channels, IEEE J. Sel. Areas Commun. 25 (2007) 10911100.

Song Han is currently working toward the Ph.D.

degree with the Key Lab of Industrial Computer
Control Engineering of Hebei Province, Yanshan
University, Qinhuangdao, China. His research interests include resource allocation in femtocell
networks.

Xinbin Li received his M.Sc. degree in Control

Theory and Control Engineering from Yanshan
University, China in 1999, and the Ph.D. degree in
General and Fundamental Mechanics from Peking
University, China in 2004. He is now a professor
in the Institute of Electrical Engineering, Yanshan
University, China. His research interests include
nonlinear system and femtocell networks.

Zhixin Liu received his B.S., M.S., and Ph.D. degrees in Control Theory and Engineering from
Yanshan University, Qinhuangdao, China, in 2000,
2003, and 2006, respectively. He is currently a
professor with the Department of Automation,
Institute of Electrical Engineering, Yanshan University, China. He visited the University of Alberta, Edmonton, AB, Canada, in 2009. His current research interests include performance optimization and energy-ecient protocol design in
wireless sensor networks, wireless resource allocation in cognitive radio networks and femtocell
networks.
Xinping Guan received his Ph.D degree in Control
and Systems from the Harbin Institute of Technology, China. He joined the Department of Automation, Shanghai Jiao Tong University, China,
in 2007 where he is currently the Distinguished
University Professor, Director of Key Laboratory
of Systems Control and Information Processing,
Ministry of Education of China. Prof. Guan has
authored and/or coauthored two research monographs, more than 120 SCI indexed papers in IEEE
Transactions and other well-known international
journals and numerous conference papers. His
current research interests include wireless sensor
networks, ground-air communication of aircrafts, and cognitive radio networks and their applications in industry. As principle investigator, Prof. Guan
has nished/been working on many national key projects. He is the leader
of the prestigious Innovative Research Team of NSFC in 2012. He was one
of the rst prize winners of the University Natural Science Award from the
Ministry of Education of China in 2003 and 2007, respectively. He received
the IEEE Transaction on Fuzzy Systems Outstanding Paper Award in 2008
and also received the second prize of the National Natural Science Award of
China. Prof. Guan was appointed as Changjiang Scholar by the Ministry
of Education of China, and State-level Scholar of New Century Bai Qianwan Talent Program of China. He received the National Outstanding Youth
Award from the National Natural Science Foundation of China (NSFC). He
is also a Cyber Principle Investigator with the Cyber Joint Innovation Center
founded by Zhejiang University, Tsinghua University, and Shanghai Jiao Tong
University. Prof. Guan is a senior member of IEEE. He is currently a committee member of the Chinese Automation Association Council and Chinese
Articial Intelligence Association Council. He serves as an Associate Editor
for IEEE Transaction on System, Man and Cybernetics-C, as Editorial Board
Committee Member for the several Chinese journals, and an International
Technical Committee Member for a lot of conferences. Wirel.